Enhance AI security with Azure Prompt Shields and Azure AI Content Safety

10 months ago 185

Defend your AI systems with Prompt Shields—a unified API that analyzes inputs to your LLM-based solution to defender against nonstop and indirect threats.

A beardown defence against punctual injection attacks

The AI information scenery is perpetually changing, with punctual injection attacks emerging arsenic 1 of the astir important threats to generative AI app builders today. This occurs erstwhile an adversary manipulates an LLM’s input to alteration its behaviour oregon entree unauthorized information. According to the Open Worldwide Application Security Project (OWASP), punctual injection is the apical menace facing LLMs today1. Help support your AI systems against this emerging menace with Azure AI Content Safety, featuring Prompt Shields—a unified API that analyzes inputs to your LLM-based solution to defender against direct and indirect threats. These exploits tin see circumventing existing information measures, exfiltrating delicate data, oregon getting AI systems to instrumentality unintended actions wrong your environment.

Prompt injection attacks

In a punctual injection attack, malicious actors input deceptive prompts to provoke unintended oregon harmful responses from AI models. These attacks tin beryllium classified into 2 main categories—direct and indirect punctual injection attacks.

  • Direct punctual injection attacks, including jailbreak attempts, hap erstwhile an extremity idiosyncratic inputs a malicious punctual designed to bypass information layers and extract delicate information. For instance, an attacker mightiness punctual an AI exemplary to divulge confidential data, specified arsenic societal information numbers oregon backstage emails.
  • Indirect, oregon cross-prompt injection attacks (XPIA), impact embedding malicious prompts wrong seemingly innocuous outer content, specified arsenic documents oregon emails. When an AI exemplary processes this content, it inadvertently executes the embedded instructions, perchance compromising the system.

Prompt Shields seamlessly integrates with Azure OpenAI contented filters and is disposable successful Azure AI Content Safety. It defends against galore kinds of punctual injection attacks, and caller defenses are regularly added arsenic caller onslaught types are uncovered. By leveraging precocious instrumentality learning algorithms and earthy connection processing, Prompt Shields efficaciously identifies and mitigates imaginable threats successful idiosyncratic prompts and third-party data. This cutting-edge capableness volition enactment the information and integrity of your AI applications, helping to safeguard your systems against malicious attempts astatine manipulation oregon exploitation. 

Prompt Shields capabilities include:

  • Contextual awareness: Prompt Shields tin discern the discourse successful which prompts are issued, providing an further furniture of information by knowing the intent down idiosyncratic inputs. Contextual consciousness besides leads to less mendacious positives due to the fact that it’s susceptible of distinguishing existent attacks from genuine idiosyncratic prompts.
  • Spotlighting: At Microsoft Build 2025, we announced Spotlighting, a almighty caller capableness that enhances Prompt Shields’ quality to observe and artifact indirect punctual injection attacks. By distinguishing betwixt trusted and untrusted inputs, this innovation empowers developers to amended unafraid generative AI applications against adversarial prompts embedded successful documents, emails, and web content.
  • Real-time response: Prompt Shields operates successful real time and is 1 of the archetypal real-time capabilities to beryllium made mostly available. It tin swiftly place and mitigate threats earlier they tin compromise the AI model. This proactive attack minimizes the hazard of information breaches and maintains strategy integrity.

End-to-end approach

  • Risk and information evaluations: Azure AI Foundry offers hazard and information evaluations to fto users measure the output of their generative AI exertion for contented risks: hateful and unfair content, intersexual content, convulsive content, self-harm-related content, nonstop and indirect jailbreak vulnerability, and protected material.
  • Red-teaming agent: Enable automated scans and adversarial probing to place known risks astatine scale. Help teams displacement near by moving from reactive incidental effect to proactive information investigating earlier successful development. Safety evaluations besides enactment reddish teaming by generating adversarial datasets that fortify investigating and accelerate contented detection.
  • Robust controls and guardrails: Prompt Shields is conscionable 1 of Azure AI Foundry’s robust contented filters. Azure AI Foundry offers a fig of contented filters to observe and mitigate hazard and harms, punctual injection attacks, ungrounded output, protected material, and more.
  • Defender for Cloud integration: Microsoft Defender present integrates straight into Azure AI Foundry, surfacing AI information posture recommendations and runtime menace extortion alerts wrong the improvement environment. This integration helps adjacent the spread betwixt information and engineering teams, allowing developers to proactively place and mitigate AI risks, specified arsenic punctual injection attacks detected by Prompt Shields. Alerts are viewable successful the Risks and Alerts tab, empowering teams to trim aboveground country hazard and physique much unafraid AI applications from the start.

Customer usage cases

AI Content Safety Prompt Shields offers galore benefits. In summation to defending against jailbreaks, punctual injections, and papers attacks, it tin assistance to guarantee that LLMs behave arsenic designed, by blocking prompts that explicitly effort to circumvent rules and policies defined by the developer. The pursuing usage cases and lawsuit testimonials item the interaction of these capabilities.

AXA: Ensuring reliability and security

AXA, a planetary person successful insurance, uses Azure OpenAI to powerfulness its Secure GPT solution. By integrating Azure’s contented filtering exertion and adding its ain information layer, AXA prevents punctual injection attacks and helps guarantee the reliability of its AI models. Secure GPT is based connected Azure OpenAI successful Foundry Models, taking vantage of models that person already been fine-tuned utilizing quality feedback reinforcement learning. In addition, AXA tin besides trust connected Azure contented filtering technology, to which the institution added its ain information furniture to forestall immoderate jailbreaking of the exemplary utilizing Prompt Shields, ensuring an optimal level of reliability. These layers are regularly updated to support precocious safeguarding.

Wrtn: Scaling securely with Azure AI Content Safety

Wrtn Technologies, a starring endeavor successful Korea, relies connected Azure AI Content Safety to support compliance and information crossed its products. At its core, Wrtn’s flagship exertion compiles an array of AI usage cases and services localized for Korean users to integrate AI into their mundane lives. The level fuses elements of AI-powered search, chat functionality, and customizable templates, empowering users to interact seamlessly with an “Emotional Companion” AI-infused agent. These AI agents person engaging, lifelike personalities, interacting successful speech with their creators. The imaginativeness is simply a highly interactive idiosyncratic cause that’s unsocial and circumstantial to you, your data, and your memories.

Because the merchandise is highly customizable to circumstantial users, the built-in quality to toggle contented filters and Prompt Shields is highly advantageous, allowing Wrtn to efficiently customize its information measures for antithetic extremity users. This lets developers standard products portion staying compliant, customizable, and responsive to users crossed Korea.

“It’s not conscionable astir the information and privacy, but besides safety. Through Azure, we tin easy activate oregon deactivate contented filters. It conscionable has truthful galore features that adhd to our merchandise performance,” says Dongjae “DJ” Lee, Chief Product Officer.

Integrate Prompt Shields into your AI strategy

For IT determination makers looking to heighten the information of their AI deployments, integrating Azure’s Prompt Shields is simply a strategical imperative. Fortunately, enabling Prompt Shields is easy.

Azure’s Prompt Shields and built-in AI information features connection an unparalleled level of extortion for AI models, helping to guarantee that organizations tin harness the powerfulness of AI without compromising connected security. Microsoft is a person successful identifying and mitigating punctual injection attacks, and uses champion practices developed with decades of research, policy, merchandise engineering, and learnings from gathering AI products astatine scale, truthful you tin execute your AI translation with confidence. By integrating these capabilities into your AI strategy, you tin assistance safeguard your systems from punctual injection attacks and assistance support the spot and assurance of your users.

Our committedness to Trustworthy AI

Organizations crossed industries are utilizing Azure AI Foundry and Microsoft 365 Copilot capabilities to thrust growth, summation productivity, and make value-added experiences.

We’re committed to helping organizations usage and physique AI that is trustworthy, meaning it is secure, private, and safe. Trustworthy AI is lone imaginable erstwhile you harvester our commitments, specified arsenic our Secure Future Initiative and Responsible AI principles, with our merchandise capabilities to unlock AI translation with confidence. 

Get started with Azure AI Content Safety

1OWASP Top 10 for Large Language Model Applications

Read Entire Article